Stage Prediction of Alzheimer’s Disease Based on 3D CNN

Alzheimer’s disease (AD) is a growing global health problem especially for the elders. It is affecting over 50 million people in the world. Cognitive decline and memory impairment are its hallmarks. Currently, the survival time after being diagnosed is short. Therefore, we need to make the detection as early as possible and at the same time accurate to achieve an effective intervention. Traditional diagnosis methods rely on time-consuming manual analysis and at the same time may produce error. A three-dimensional convolutional neural network (3D CNN) will be developed in this study to automatically classify AD stages from structural magnetic resonance imaging (MRI) scans. The dataset is from OASIS. The 3D CNN architecture in this research consists of four convolutional blocks with the following specifications: Wprogressive filter sizes, batch normalization, dropout, ReLU activation, 3D max pooling and global average pooling. Training results showed a steady decrease in loss and increase in accuracy. However, the training accuracy is relatively low and the model starts overfitting in a short training period. The results showed that the model can learn meaningful features from MRI data, but a larger and balanced dataset is essential to improve the generalization and achieve a reliable AD stage classification.